Deep Restoration Priors (DRP) is a new method for using pre-trained restoration models as priors for inverse problems
Image denoisers have been shown to be powerful priors for solving inverse problems in imaging. In this work, we introduce a generalization of these methods that allows any image restoration network to beused as an implicit prior. The proposed method uses priors specified by deep neural networks pre-trained as general restoration operators. The method provides a principled approach for adapting state-of-the-art restoration models for other inverse problems. Our theoretical result analyzes its convergence to a stationary point of a global functional associated with the restoration operator. Numerical results showthat the method using a super-resolution prior achieves state-of-the-art performance both quantitatively and qualitatively. Overall, this work offers a step forward for solving inverse problems by enabling the use of powerful pre-trained restoration models as priors.
Figure 1: Visual comparison of DRP with several well-known methods on Gaussian deblurring of color images. Note how DRP using restoration prior improves over SOTA methods based on denoiser priors.
Figure 2: Visual comparison of DRP and several well-known methods on single image super resolution. Note the excellent performance of the proposed DRP method using the SwinIR prior both visually and in terms of PSNR.
Figure 3: llustration of the convergence behaviour of DRP for image deblurring and single image superresolution on the Set3c dataset. (a)-(b): Deblurring with Gaussian blur kernels of standard deviations 1.6 and 2.0. (c)-(d):2× and 3× super resolution with the Gaussian blur kernel of standard deviation 2.0.
Figure 4: Illustration of the impact of different SR factors in the prior used within DRP for image deblurring. We show three scenarios: (i) using only 3× prior, (ii) using only 2× prior, and (iii) the use of the prior refinement strategy, which combines both the 2× and 3× priors. Left: Convergence of PSNR against the iteration number for all three configurations. Right: Visual illustration of the final image for each setting. The visual difference is highlighted by the red arrow in the zoom-in box. Note how the reduction of q can lead to about 0.3 dB improvement in the final performance.
Figure 6: llustration of denoising results of DRP on CBSD68 dataset with noise level σ= 0.1. Each image is labeled by its PSNR (dB) with respect to the original image.